Who's holding the steering wheel?

As software starts to act on our behalf, the most important thing a designer can protect is the user's sense of being in control. Agency isn't a feature you ship, it's the difference between a tool and a trap.

There's a quiet shift happening in how we build software. For two decades, the job of design was to help people do things - click, type, drag, navigate. Now we're building systems that do things for people. Autocomplete became autodraft. Recommendation became decision. The interface stopped being a set of controls and started becoming a collaborator with opinions of its own.

That shift is wonderful and it is dangerous, often in the same breath. Every bit of work a system takes off a person's plate is also a bit of control quietly handed over. The central design question of this era isn't "can the system do it?" It almost always can. It's "does the person still feel like they're the one driving?"

That feeling has a name in human-computer interaction: the sense of agency. It's the experience of being the author of your own actions of choices being yours, outcomes being traceable to your intent. Decades of research show it's not a nice-to-have. When agency drops, so does trust, satisfaction, and the willingness to use a product at all. And the cruel twist is that automation, the very thing meant to help, is the fastest way to erode it.

Look at the curve and the design brief writes itself. Strip away every bit of automation and you get raw control but exhausting friction, people give up. Automate everything and you get effortless output that no longer feels like the person's own, people disengage, or worse, stop noticing what the system decided for them. The job is to design for the middle: enough help to be magical, enough control to stay yours.

THE FACTORS

Six things to get right

Designing for agency isn't abstract. It comes down to a handful of concrete decisions you make on every screen, every flow, every model output. Here's the checklist I keep coming back to.

01

Meaningful control, not just buttons

Surface the steering wheel. People should be able to redirect a system at the moments that matter, not just accept or reject a finished result. Control that arrives only after the work is done isn't control.

02

Transparency over black boxes

When a system makes a choice, the person should be able to see enough of why to trust it or challenge it. Opacity feels efficient right up until something goes wrong and nobody can explain it.

03

Reversibility as a default

Agency thrives where experimentation is safe. If every action has a clean undo, people explore boldly. If mistakes are permanent, they freeze, and a frozen user has no real freedom.

04

Honest defaults

Most people never change a default, which makes defaults the most powerful design decision you'll make. Set them for the user's benefit, not for the metric you're trying to move this quarter.

05

Keep the human in the loop

Automation should scaffold, not substitute. Show editable, intermediate steps, a draft, a wireframe, a stem, so the person stays an author rather than a spectator to their own work.

06

Ownership and attribution

In creative software especially, agency includes a question older than computing: who owns what comes out, and who gets the credit? A system that obscures provenance has taken something more than effort.

The measure of good AI design isn't how much it does for you. It's how much it leaves you in charge of. A principle we build on at Soundverse AI

MY VIEW

Why I think agency is the whole game

I spent years building creation tools at Spotify and the music-making side, at Samsung on voice. The pattern I kept seeing is that the tools people love aren't the ones that do the most. They're the ones that make people feel more capable than they were the day before. That feeling is agency, and it's fragile. One over-eager automation, one moment where the software decided something you didn't ask for, and the relationship tips from "my instrument" to "someone else's machine."

Generative AI has made this urgent. We can now produce a finished song, image, or document from a single sentence. That's genuinely magical and it's exactly the point on the curve where agency falls off a cliff. If a creator types a prompt and a complete track falls out, what did they actually make? Who made it? What was it trained on? If the tool can't answer those questions, it hasn't empowered the creator. It's replaced them and handed back the receipt.

So I've come to believe the differentiator in AI products won't be raw generation quality that's commoditizing fast. It'll be agency architecture: how deliberately a product keeps the human authoring, steering, owning, and getting paid. That belief is the spine of how we build Soundverse.

IN PRACTICE

How this shows up at Soundverse

Soundverse is an AI music platform, and music is the perfect stress test for agency because a song is personal, ownership is contested, and the temptation to just hand someone a finished WAV file is enormous. Most AI music tools stop exactly there: prompt in, track out, no idea what it was trained on or what you're allowed to do with it. We made a deliberate decision to design the opposite way. Three parts of our stack exist specifically to protect agency.

Agent One: a co-producer, not a vending machine

Agent One is a conversational AI that compresses what used to take a whole label team, creating, owning, marketing, distributing into a single agent you talk to. But the design principle underneath it is that the creator stays the director. The agent proposes; you decide. It carries you across the lifecycle from idea to audience without ever taking your hands off the wheel. That's points 1 and 5 of the framework, made concrete.

Artist DNA: agency for the artists too

Agency can't only protect the user in front of the screen. The artists whose style trains a model have agency at stake as well. Artist DNA is a marketplace of permissioned, artist-owned AI models trained only with consent, on terms the artist sets. Nobody's voice or signature gets used because it was scrapeable. That's honest defaults and ownership applied to the people most AI systems forget.

TRACE: making provenance visible

TRACE is our attribution and fair-pay layer: watermarking, data attribution, and licensing tags that travel with the work. It's the answer to "what went into this, and who should get paid?" the transparency and ownership factors turned into infrastructure. When a creator can see the lineage of what they made, they're not just using a tool. They're standing behind their work.

The takeaway

The next decade of software will be defined by systems that act for us. The good ones will be the ones that never let us forget we're the author. Design for agency, and you don't just build a more ethical product, you build one people trust enough to make their own.

Agency is not a constraint on what AI can do for people. It's the thing that makes what AI does worth doing. Build the steering wheel first. Everything else is just horsepower.